You are on page 1of 16

Prediction of Angular Distortion in Gas Tungsten Arc Welded 202

Grade Stainless Steel Plates Using Artificial Neural Networks An


Experimental Approach
Sudhakaran .R
a
, VeL Murugan .V
b
, Siva Sakthivel. P. S
c
, Sakthivel .P
d
Sudhakaran .R
a
(Corresponding author)
Senior Lecturer
Department of Mechanical Engineering
Kumaraguru College of Technology
Coimbatore 641006, Tamil Nadu, India.
Ph : 91 422 2669401 (O), Fax: 91 422 2669406
Mobile: 91 9894030121
E mail : absudha@yahoo.com
Dr. VeL Murugan. V
b

Prof and HOD
Department of Aeronautical Engineering
Kumaraguru College of Technology
Coimbatore 641006, Tamil Nadu, India
Siva Sakthivel. P. S
c
Senior Lecturer
Department of Mechanical Engineering
Kumaraguru College of Technology
Coimbatore 641006, Tamil Nadu, India
Sakthivel .P
d
Faculty
Department of Mechanical Engineering
Kumaraguru College of Technology
Coimbatore 641006, Tamil Nadu, India.
Suggested Reviewers
Dr. N. Alaghumoorthy
Professor
Department of Mechanical Engineering
Pondicherry Engineering College
Pondicherry, India
e mail: alagu_pec@yahoo.co.in
Dr. N. Gunaraj
Principal
RVS College of Engineering
Coimbatore, India
e mail: vgunadeepakkct@yahoo.com
Dr. A. Ragupathy
Professor
Department of Mechanical Engineering
Annamalai University
Annamalai Nagar 608002
India
e mail: aragupathy_au@hotmail.com
Prediction of Angular Distortion in Gas Tungsten Arc Welded 202 Grade
Stainless Steel Plates Using Artificial Neural Networks An Experimental
Approach
Sudhakaran. R
a
, Vel Murugan. V
b
, Siva Sakthivel. P. S
c
, Sakthivel .P
d

a
Department of Mechanical Engineering, Kumaraguru College of Technology,
Coimbatore 641006, Tamilnadu, India, absudha@yahoo.com
b
Department of Aeronautical Engineering, Kumaraguru College of Technology,
Coimbatore 641006, Tamilnadu, India, velgodham_v@yahoo.com
c
Department of Mechanical Engineering, Kumaraguru College of Technology,
Coimbatore 641006, Tamilnadu, India
d
Department of Mechanical Engineering, Kumaraguru College of Technology,
Coimbatore 641006, Tamilnadu, India
Abstract
This paper presents development of an Artificial Neural Network (ANN) model to
predict angular distortion for various input process parameters in 202 grade stainless
steel gas tungsten arc welded plates. The chosen input process parameters were
welding gun angle, welding current, plate thickness, welding speed and gas flow rate.
The chosen output parameter is angular distortion. The experiments were conducted
based on five factor five level central composite rotatable designs with full replication
technique. Using the experimental data multi layer feed forward neural network
model was developed and it was trained by using back propagation algorithm. The
developed model is then compared with the experimental results and it is found that
the result obtained from neural network model is accurate in predicting the angular
distortion.
Keywords
Welding parameters, Angular distortion, Back propagation, Stainless steel, Regression
analysis
1.0 Introduction
Gas Tungsten Arc Welding is an arc welding process that produces coalescence of
metals by heating them with an arc between a non consumable electrode and base
metal. GTAW process is suitable for joining thin and medium thickness materials like
stainless steel sheets and for applications where metallurgical control of the weld
metal is critical. Stainless steel 202 grade has wide applications in making seamless
stainless steel tubes for boilers, heat exchangers tubes, super heater tubes, cook wares
etc., Angular distortion or out of plane distortion is one such defect that makes the
work piece distort in angular directions around the weld interface. Post weld treatment
is required to eliminate the distortion so that the work piece is defect free and
accepted. The extent of angular distortion is directly influenced by the welding input
parameters during the welding process; therefore, welding can be considered as a
multi input process [1]. One of the methods to remove the angular distortion during
the fabrication process is to provide an initial angular distortion in the negative
direction. If an exact magnitude of angular distortion is predicted, then a weld with no
angular distortion would be the result. It is difficult to obtain analytical solution to
predict angular distortion. Costly and time consuming experiments are required in
order to determine the optimum welding process parameters due to the complex and
linear nature of the welding process. Therefore, a more efficient method is needed to
determine the optimum welding process parameters. The technique of neural
networks offers potential as an alternative to standard computer techniques in control
technology and has attracted a widening interest in their development and application
[2]. The advantage of neural networks is that the network can be updated continuously
with new data to optimize its performance at any instant, the networks ability to
handle a large number of input variables rapidly and the networks ability to filter
noisy data and interpolate incomplete data [2]. Watanabe and Sastoh [3] used a
combination of empirical and analytical methods to study the effects of welding
conditions on the distortion in the welded structures. Mandal and Parmar [4] used a
statistical method of two level full factorial techniques to develop mathematical
model and reported that welding speed had a positive effect on angular distortion.
Gunaraj and VeL Murugan and [5] studied the effect of process parameters on angular
distortion in Gas Metal Arc Welding (GMAW) of structural steel plates. They
developed a mathematical model for angular distortion based on five level five
factorial central composite designs. Li et al [6] have proposed a neural network for on
line prediction of quality in GMAW. They showed that accurate quality prediction
was possible for both short circuiting and spray metal transfer modes. Kim et al [7]
compared multiple regression and back propagation neural network approaches in
modeling top bead height of multi pass GMAW. They reported that back propagation
neural network was considerably more accurate than multiple regression techniques.
Nagesh and Dutta [8] applied a back propagation neural network to predict weld bead
geometry and penetration in Shielded Metal Arc Welding (SMAW). They reported
that artificial neural networks are powerful tool for the analysis and modeling of weld
bead geometry and penetration. Control of distortion in Robotic CO
2
shielded flux
cored arc welding was investigated by Arya and Parmar [9]. A three level fractional
factorial technique was used to predict the angular distortion in 10 mm thick low
carbon steel. The effect of arc voltage, wire feed rate, welding speed and groove angle
on the angular distortion in single vee butt welds was investigated with and without
sealing run. It was concluded that the models developed were fairly accurate and can
be usefully employed for controlling the angular distortion in automated welding lines
using the flux cored arc welding process. The prediction of laser butt joint parameters
using back propagation and learning vector quantization networks was done by Jeng
et al. [10]. The work piece thickness and welding gap was used as input parameters.
The optimal focused position, acceptable welding parameters of laser power and
welding speed, and welding quality were used as output parameters. They concluded
that both networks are very useful in selecting welding parameters. Tarang et al. [11]
utilized back propagation neural network to study the relationships between the
process parameters and the features of the bead geometry for the GTA welding.
Kannan and Murugan [12] developed a neural network model for predicting the weld
bead geometry and dilution in flux cored arc welding. They found that the
performance of the model was very accurate in predicting the weld bead geometry
and dilution. It was confirmed that artificial neural networks are powerful tools for
analysis, modeling and control of such applications. The results suggested that
artificial neural networks can yield real time results of equal or better accuracy and
reliability than other data analysis algorithms. A lot of work has been done in the area
of analyzing the weld bead geometry using ANN. However, there is very little
published information available with regard to modeling and prediction of angular
distortion in 202 grade stainless steel GTAW plates. This paper presents development
of a neural network model to predict angular distortion in 202 grade stainless steel
GTAW plates. The chosen input parameters were welding gun angle, welding speed
and welding current, gas flow rate and plate length. The output parameter is angular
distortion. The experiments were conducted based on five factor five level central
composite rotatable design with full replications technique. Using the experimental
data multi layer feed forward neural network model was developed and it was trained
using back propagation algorithm. The developed model is then compared with the
experimental results and it is found that the result obtained from neural network
model is accurate in predicting the angular distortion.
2. Experimental Procedure
The experiments were conducted using Lincoln V 350 Pro electric digital welding
machine. The experimental set up consists of a traveling carriage with a table for
supporting the specimens. The welding torch was held stationary in a frame mounted
above the work table and it was provided with an attachment for both up and down
movement and angular movement for setting the required welding gun angle. A servo
motor driven manipulator was used to maintain uniform welding speed. Test plates of
following sizes
(100mmX35mmX3mm), (125mmX35mmX3mm), (150mmX35mmX3mm),
(175mmX35mmX3mm) and
(200mmX35mmX3mm) are cut from grade 202 stainless steel plates and one surface
was cleaned to remove oxide scale and dirt before welding. Argon is used as the
shielding gas and its flow rate is varied for each experiment as per the requirements.
3. Plan of Investigation
The research was carried out in the following steps
1. Identification of the process parameters.
2. Finding the limits of the process parameters
3. Developing the design matrix
4. Conducting the experiments
5. Recording the responses i.e., angular distortion
6. Developing a neural network model
7. Testing the network
8. Performing simulation.
3.1 Identification of the process parameters
The extent of angular distortion highly depends on the welding process control
parameters. The following independently controllable process parameters were
identified to carry out the experiments: welding current (I), welding speed (V), gas
flow rate (Q), gun angle (), and plate length (L).
3.2 Finding the limits of the process parameters
The working ranges of all selected factors are fixed by conducting trial runs. This was
carried out by varying one of the factors while keeping the rest of them as constant
values. The working range was decided upon by inspecting the bead for a smooth
appearance without any visible defects such as surface porosity, under cut etc., The
upper limit of a given factor was coded as +2 and the lower limit was coded as -2. The
coded values for intermediate values were calculated using the Eq. (1)
min
X
max
X
))
min
X
max
(X 2(2X
i
X

(1)
Where X
i
is the required coded value of a variable X and is any value of the variable
from X
min
to X
max
. The selected process parameters with their limits and notations are
given in Table 1.
3.3 Developing the design matrix
In factorial design, the experiments are conducted for all possible combinations of the
parameter levels and these combinations written in the form of a table where the rows
correspond to different trials and the columns to the levels of the parameters, form a
design matrix. The design matrix selected for experiment is a five factor five level
central composite rotatable design with 32 experimental runs [14]. The experiments
were conducted as per the design matrix at random, to avoid the possibility of
systematic errors infiltrating the system. The five process parameters , V, I, L and Q
are set as per the design matrix and the experiments are conducted.
3.4 Recording the response Angular Distortion
The angular distortion was measured using Microscribe G2 coordinate measuring
machine. The work piece under observation is clamped to the desk with the help of C-
clamp. With the help of trisquare, straight lines are drawn to scribe at three or four
places in the work piece. The Microscribe G2 is interfaced with the Rhino 4 software.
The limits for the work piece in the X and Y directions are fixed i.e. boundary for
indicating the domain of the work piece. The angle between the two lines is
measured as shown in the Fig.1.
The distorted angle can be obtained by the relation given by Eq. (2).
2 ) (180
(2)
Four readings were taken randomly on each welded plate and the average value is
recorded. The Table.2 shows the design matrix, the measured value of angular
distortion, the predicted results and percentage error.
3.5 Developing a neural network model
Artificial neural networks are one of the most powerful computer modeling
techniques based on statistical approach, currently being used in many fields of
engineering for modeling complex relationships which are difficult to describe with
physical models. The attraction of neural networks comes from their remarkable
information, processing characteristics pertinent mainly to non linearity, high
parallelism, fault and noise tolerance, and learning and generalized capability. There
has been continual increase in research interest in the application of artificial neural
networks in modeling and monitoring of welding processes. The objective of this
study is to model the angular distortion of 202 grade stainless steel plates in GTAW
process. The five steps used in general application of neural network are
1. Collection of input/output data set
2. Pre processing of input/ output data set
3. Neural network design and training
4. Testing of the network
5. Performing simulation
3.5.1 Collection of input/output data set
The input parameters used to train the network consists of 5 neurons namely welding
current (I), welding speed (V), gas flow rate (Q), gun angle (), and plate length (L).
The output parameter used to train the network consists of one neuron namely angular
distortion. The input and output data set used for training the network is shown in
Table 2.
3.5.2 Preprocessing of input/output dataset
The generalization capability of the neural network is essentially dependent on
(i) The selection of appropriate input/output parameters of the process.
(ii) The distribution of the dataset.
(iii) The format of the presentation of the data set to the network.
In total 27 experimental data were collected for building the neural network model.
Among these 21 data were selected as training data. The residuals (i.e.6 data) were
used to verify the predicted values. This is shown in Table .2. In order to relieve the
training difficulty and balance the importance of each parameter during training
process, the data should be normalized. It is recommended that the data be normalized
between slightly offset values such as 0.1 and 0.9 rather than between 0 and 1 to avoid
saturation of sigmoid function leading to slow or no learning. The normalized value
for each raw input/output dataset was calculated using the following equation

,
_

+
min
X
max
X
min
X X
))
1

2
(
1
(
'
X
(3)
Where X
1
= Normalized value of X

1
= 0.1,
2
= 0.9
X
max
= maximum value of X
X
min
= minimum value of X
3.5.3 Neural network design and training
The performance of the neural network depends on the number of hidden layers and
number of neurons in the hidden layers. Therefore many attempts should be carried
out in choosing the optimal structure for the neural network by changing the number
of hidden layers as well as the number of neurons in each of these hidden layers. The
appropriate neural network structure for predicting angular distortion was chosen by
trail error method. In this study, a feed forward back propagation artificial neural
network was created keeping five neurons in the input layer, one neuron in the hidden
layer and one neuron in the output layer using MATLAB 7.6. The number of neurons
in the hidden layer was varied between 1 and 25. The structure of the designed neural
network was 5 5 1. The structure is shown in Fig. 2
3.5.4 Testing the network
The network was tested to determine the performance of established model of angular
distortion. The network was trained for 14 iterations. Further training did not improve
the performance of the network. The average error between the desired and actual
output is less than 0.00014 at the end of the training process. This is shown in Fig.4.
The performance of the model was carried on the testing data set. The network model
is shown in Fig.3. The percentage of error of the model predicted was calculated as
the percentage difference between the experimental and predicted value relative to the
experimental value. The results show that the percentage error is less than 5.5%. This
is shown in Table 2. The regression analysis is performed to find out the correlation
coefficient. The correlation coefficient is used to measure the relationship between the
measured and predicted values. An R value of 1 means a close relationship and 0 a
random relationship. It is observed that a regression coefficient of R = 0.99691 is
obtained for training data, R = 0.99536 for testing data and R = 0.93851 for all the
data. This demonstrated that the model has a high accuracy for predicting the angular
distortion. The line of best fit was drawn using the plotted points and is calculated
using the regression analysis. This is shown from Fig.5 to Fig.7.
3.5.5 Performing simulation
The influence of process parameters on the angular distortion was studied using the
model. The angular distortion was predicted by varying one of the parameters and
keeping the remaining parameters constant. This is done by using the developed
neural network model and the predicted values are represented in graphical form from
Figures 8 to 12.
4. Results and discussion
The effect of welding parameters such as welding gun angle, welding speed, welding
current, gas flow rate and plate length on angular distortion are discussed below.
4.1 Effect of welding gun angle on angular distortion
The Fig. 8 represents the effect of gun angle on angular distortion. From the figure it
is clear that the angular distortion decreases for lower gun angles. At lower gun angles
the preheating of work piece is less. This results in less depth of penetration and width
of the bead. Higher gun angles results in more penetration and width of the bead. An
increase in width of the bead will contract more at the top surface of the weld pool.
Also at higher gun angles weld metal gets more exposure to the arc which increases
the thermal stresses in the heat affected zone. Hence there is an increase in angular
distortion.
4.2 Effect of welding speed on angular distortion
Artem Pilipenko [15] reported a relationship for angular distortion
2
Sh
IV
0.13
(4)
Where I is current in Amps, V is voltage, S is welding speed in m/s and h is plate
thickness in m. Watanabe and Satoh [3] reported another relationship for angular
distortion
1
1
1
]
1

,
_

+
1
]
1

Sh h
1
2
C
e
1 m
Sh h
1
1
c
(5)
From these two equations, it is clear that the increase in welding speed results in
decrease in angular distortion. Welding speed is one of the main factor controlling
heat input and the bead width. The bead width and dimensions of heat affected zone
decrease with the increase in welding speed. This is because heat input is inversely
proportional to welding speed. As width of the bead and heat affected zone decreases
with the increase in welding speed, the angular distortion also decreases with the
increase in welding speed. This is clearly shown in Fig.9.
4.3 Effect of plate length on angular distortion
From the Fig. 10 it is clear that the angular distortion decreases with the increase in
plate length. When the plate length increases the plate gets stiffened. This resists
distortion. Also when the plate length increases there is more length available for the
thermal stresses in the heat affected zone to get distributed. Hence the cumulative
effects of the above two factors results in decrease in distortion for the corresponding
increase in plate length.
4.4 Effect of welding current on angular distortion
From Artem Pilipenko [15] the heat input is directly proportional to angular
distortion. When the welding current increases the heat input increases. The increase
in heat input results in preheating of the work piece during the forward welding. This
also results in more penetration and width of the bead. The increase in welding
current also increases the thermal stresses in the heat affected zone. The cumulative
affect of the above two factors results in increase in angular distortion for the
corresponding increase in welding current. This is shown in Fig. 11
4.5 Effect of gas flow rate on angular distortion
The Fig. 12 shows the effect of gas flow rate on angular distortion. When the gas flow
is varied from the lower level to higher level there is a decreasing trend in angular
distortion. This is because at higher gas flow rates more heat is carried by the gas.
This results in less depth of penetration and decrease in dimensions of the heat
affected zone. Hence there is a decrease in angular distortion for corresponding
increase in gas flow rate.
Conclusions
The effect of process parameters on angular distortion in GTAW 202 grade stainless
steel plates using neural networks has been stainless steel plates using neural networks
have been studied, and the following conclusions have been reached.
1. The neural network model developed in this work from the experimental data
can be employed to control the process in order to achieve the desired weld
quality in butt welded plates.
2. It is observed from the results that a correlation coefficient of 0.99 is obtained
between the results obtained by the experimental and the model developed.
Hence there exists a close relationship between the experimental and the
developed model.
3. The percentage of error obtained for angular distortion between predicted and
experimental value falls within the limit of 95% confidence level. Hence the
developed neural network model is capable of making the prediction of
angular distortion with reasonable accuracy.
4. The maximum angular distortion is 12 when the process parameters such as
welding current, gun angle, gas flow rate and plate length are maintained at 80
amps, 60 and 10 litre/min, 125 mm respectively and welding speed is
maintained at 110 mm/min.
5. The minimum angular distortion is 0.45 when gun angle, welding speed are
maintained at 60 and 90 mm/min respectively and the other process
parameters such as plate length, welding current and gas flow rate are
maintained at 175 mm, 100 amps and 20 litre/min respectively.
6. Out of the five process parameters selected for investigation, welding current
has strong effect on angular distortion; plate length, welding speed and gas
flow rate has a negative effect on angular distortion.
Acknowledgement
The author wishes to thank All India Council for Technical Education, New Delhi,
India for sponsoring this research project.
References
[1] Vinokurov V A. Welding stresses and distortion. Wetherby. British Library;
2002, p. 35 -40.
[2] Manikya Kanti K, Srinivasa Rao P. Prediction of bead geometry in pulsed GMA
welding using back propagation neural network. Journal of Materials Processing
Technology. 2008; 200: p. 300 305.
[3] Watanabe M, Sastoh K. Effect of welding conditions on the shrinkage and
distortion in welded structures. Welding Journal.1961; p. 377-s - 384- s.
[4] Mandal A, Parmar RS. Effect of process variables and angular distortion of pulse
GMAW welded HSLA plates. Indian Welding Journal. 1997; 5: p. 26 34.
[5] Gunaraj V, VeL Murugan V. Effect of process parameters on angular distortion of
gas metal arc welded structural steel plates. Welding Journal.2005; p. 165 -s -
171-s
[6] Li X, Simpson SW, Rados M. Neural networks for online prediction of quality in
gas metal arc welding. Science and Technology of Welding and Joining. 2000;
5(2):p. 71 79.
[7] Kim IS, Lee SH, Yarlagadda PKDV. Comparison of multiple regression and back
propagation neural network approaches in modeling top bead height of multipass
gas metal arc welds. Science and Technology of Welding and Joining. 2003; 8(5):
p. 347 352.
[8] Nagesh DS, Datta GL. Prediction of weld bead geometry and penetration in
shielded arc welding using artificial neural networks. Journal of Materials
Processing Technology. 2002; 123(2): P. 303 312.
[9] Arya S, Parmar RS. Mathematical models for predicting angular distortions in
CO
2
shielded flux cored arc welding. Proceedings of the International
Conference on Joining of Metals (JOM 3), Helsingor, Denmark. 1986; p. 240
245.
[10] Jeng Ywan Jeng, Tzuoh Fei Man, Shyeu Ming Leu. Prediction of laser butt
joints welding parameters using back propagation and learning vector quantization
networks. Journal of Materials Processing Technology. 2000; 99(2): p. 207 218.
[11] Tarang YS, Tsai HL, Ych SS. Modeling, optimization and classification of weld
quality in tungsten inert gas welding. International Journal of Machine Tools and
Manufacture. 1999; 39: p. 1427 1438.
[12] Kannan T, Murugan N. Artificial neural network modeling of weld bead
geometry and dilution in flux cored arc welding. International Journal for Joining
of Materials.2006; 18(2): p. 45 52.
[13] DOE PC IV. Software Reference Manual. Quality America, Inc; 1992, p. 17
46.
[14] Cochran WG, Cox GM. Experimental Designs. Asia Publishing House; 1963,
p.334 374
[15] Pilipenko A. Computer simulation of residual stresses and distortion of thick
plates in multi electrode submerged arc welding their mitigation techniques.
PhD thesis; Norwegian University of Science and Technology: Norway. p.40
82.
List of Tables and Figure captions
Table 1 Welding Parameters and their levels
Table 2 Design matrix and response
Fig. 1 Angular Distortion Measured from Microscribe G2 Coordinate Measuring
Machine
Fig. 2 Feed forward neural network architecture for predicting angular distortion
Fig. 3 Network model obtained from MATLAB
Fig. 4 Performance goal of the neural network model obtained from MATLAB
Fig. 5 Line of best fit for predicted and actual angular distortion by BPN model for
training samples
Fig. 6 Line of best fit for predicted and actual angular distortion by BPN model for
testing samples
Fig. 7 Line of best fit for predicted and actual angular distortion by BPN model for all
the samples
Fig. 8 Effect of gun angle on angular distortion
Fig. 9 Effect of welding speed on angular distortion
Fig. 10 Effect of plate length on angular distortion
Fig. 11 Effect of welding current on angular distortion
Fig. 12 Effect of gas flow rate on angular distortion